Improved CPU/GPU interoperability (rapidsai#5001)

Authors:
  - Victor Lafargue (https://github.com/viclafargue)
  - William Hicks (https://github.com/wphicks)
  - Corey J. Nolet (https://github.com/cjnolet)
  - Dante Gama Dessavre (https://github.com/dantegd)
  - Carl Simon Adorf (https://github.com/csadorf)

Approvers:
  - Dante Gama Dessavre (https://github.com/dantegd)

URL: rapidsai#5001

docs/source/api.rst

@@ -180,11 +180,6 @@ Dataset Generation (Dask-based Multi-GPU)
   .. automodule:: cuml.dask.datasets.regression
      :members:
 
-Array Wrappers (Internal API)
------------------------------
-
-.. autoclass:: cuml.common.CumlArray
-    :members:
 
 Metrics (regression, classification, and distance)
 --------------------------------------------------

python/cuml/decomposition/pca.pyx

@@ -49,6 +49,7 @@ from cuml.common.exceptions import NotFittedError
 from cuml.internals.mixins import FMajorInputTagMixin
 from cuml.internals.mixins import SparseInputTagMixin
 from cuml.internals.api_decorators import device_interop_preparation
+from cuml.internals.api_decorators import enable_device_interop
 
 
 cdef extern from "cuml/decomposition/pca.hpp" namespace "ML":
@@ -406,7 +407,8 @@ class PCA(UniversalBase,
         return self
 
     @generate_docstring(X='dense_sparse')
-    def _fit(self, X, y=None) -> "PCA":
+    @enable_device_interop
+    def fit(self, X, y=None) -> "PCA":
         """
         Fit the model with X. y is currently ignored.
 
@@ -496,7 +498,8 @@ class PCA(UniversalBase,
                                        'description': 'Transformed values',
                                        'shape': '(n_samples, n_components)'})
     @cuml.internals.api_base_return_array_skipall
-    def _fit_transform(self, X, y=None) -> CumlArray:
+    @enable_device_interop
+    def fit_transform(self, X, y=None) -> CumlArray:
         """
         Fit the model with X and apply the dimensionality reduction on X.
 
@@ -541,8 +544,9 @@ class PCA(UniversalBase,
                                        'type': 'dense_sparse',
                                        'description': 'Transformed values',
                                        'shape': '(n_samples, n_features)'})
-    def _inverse_transform(self, X, convert_dtype=False,
-                           return_sparse=False, sparse_tol=1e-10) -> CumlArray:
+    @enable_device_interop
+    def inverse_transform(self, X, convert_dtype=False,
+                          return_sparse=False, sparse_tol=1e-10) -> CumlArray:
         """
         Transform data back to its original space.
 
@@ -642,7 +646,8 @@ class PCA(UniversalBase,
                                        'type': 'dense_sparse',
                                        'description': 'Transformed values',
                                        'shape': '(n_samples, n_components)'})
-    def _transform(self, X, convert_dtype=False) -> CumlArray:
+    @enable_device_interop
+    def transform(self, X, convert_dtype=False) -> CumlArray:
         """
         Apply dimensionality reduction to X.
 

python/cuml/decomposition/tsvd.pyx

@@ -35,6 +35,7 @@ from cuml.common.array_descriptor import CumlArrayDescriptor
 from cuml.common.doc_utils import generate_docstring
 from cuml.internals.mixins import FMajorInputTagMixin
 from cuml.internals.api_decorators import device_interop_preparation
+from cuml.internals.api_decorators import enable_device_interop
 
 from cython.operator cimport dereference as deref
 
@@ -299,7 +300,8 @@ class TruncatedSVD(UniversalBase,
                                                 dtype=self.dtype)
 
     @generate_docstring()
-    def _fit(self, X, y=None) -> "TruncatedSVD":
+    @enable_device_interop
+    def fit(self, X, y=None) -> "TruncatedSVD":
         """
         Fit LSI model on training cudf DataFrame X. y is currently ignored.
 
@@ -313,7 +315,8 @@ class TruncatedSVD(UniversalBase,
                                        'type': 'dense',
                                        'description': 'Reduced version of X',
                                        'shape': '(n_samples, n_components)'})
-    def _fit_transform(self, X, y=None) -> CumlArray:
+    @enable_device_interop
+    def fit_transform(self, X, y=None) -> CumlArray:
         """
         Fit LSI model to X and perform dimensionality reduction on X.
         y is currently ignored.
@@ -377,7 +380,8 @@ class TruncatedSVD(UniversalBase,
                                        'type': 'dense',
                                        'description': 'X in original space',
                                        'shape': '(n_samples, n_features)'})
-    def _inverse_transform(self, X, convert_dtype=False) -> CumlArray:
+    @enable_device_interop
+    def inverse_transform(self, X, convert_dtype=False) -> CumlArray:
         """
         Transform X back to its original space.
         Returns X_original whose transform would be X.
@@ -426,7 +430,8 @@ class TruncatedSVD(UniversalBase,
                                        'type': 'dense',
                                        'description': 'Reduced version of X',
                                        'shape': '(n_samples, n_components)'})
-    def _transform(self, X, convert_dtype=False) -> CumlArray:
+    @enable_device_interop
+    def transform(self, X, convert_dtype=False) -> CumlArray:
         """
         Perform dimensionality reduction on X.
 

python/cuml/internals/api_decorators.py

@@ -20,7 +20,6 @@
 import typing
 from functools import wraps
 import warnings
-from importlib import import_module
 
 import cuml.internals.array
 import cuml.internals.array_sparse
@@ -779,38 +778,17 @@ def inner_f(*args, **kwargs):
 
 def device_interop_preparation(init_func):
     """
-    This function serves as a decorator to cuML estimators that implement
-    the CPU/GPU interoperability feature. It imports the joint CPU estimator
-    and processes the hyperparameters.
+    This function serves as a decorator for cuML estimators that implement
+    the CPU/GPU interoperability feature. It processes the estimator's
+    hyperparameters by saving them and filtering them for GPU execution.
     """
 
     @functools.wraps(init_func)
     def processor(self, *args, **kwargs):
-        # if child class (parent class was already decorated), skip
-        if hasattr(self, '_cpu_model_class'):
-            return init_func(self, *args, **kwargs)
-
-        if hasattr(self, '_cpu_estimator_import_path'):
-            # if import path differs from the one of sklearn
-            # look for _cpu_estimator_import_path
-            estimator_path = self._cpu_estimator_import_path.split('.')
-            model_path = '.'.join(estimator_path[:-1])
-            model_name = estimator_path[-1]
-        else:
-            # import from similar path to the current estimator
-            # class
-            model_path = 'sklearn' + self.__class__.__module__[4:]
-            model_name = self.__class__.__name__
-        self._cpu_model_class = getattr(import_module(model_path), model_name)
-
         # Save all kwargs
         self._full_kwargs = kwargs
         # Generate list of available cuML hyperparameters
         gpu_hyperparams = list(inspect.signature(init_func).parameters.keys())
-        # Save list of available CPU estimator hyperparameters
-        self._cpu_hyperparams = list(
-            inspect.signature(self._cpu_model_class.__init__).parameters.keys()
-        )
 
         # Filter provided parameters for cuML estimator initialization
         filtered_kwargs = {}
@@ -824,3 +802,15 @@ def processor(self, *args, **kwargs):
 
         return init_func(self, *args, **filtered_kwargs)
     return processor
+
+
+def enable_device_interop(gpu_func):
+    @functools.wraps(gpu_func)
+    def dispatch(self, *args, **kwargs):
+        # check that the estimator implements CPU/GPU interoperability
+        if hasattr(self, 'dispatch_func'):
+            func_name = gpu_func.__name__
+            return self.dispatch_func(func_name, gpu_func, *args, **kwargs)
+        else:
+            return gpu_func(self, *args, **kwargs)
+    return dispatch